Manu Goyal

Manu Goyal is a founding engineer at Braintrust, building infrastructure that helps teams evaluate, monitor, and improve AI applications before unreliable behavior reaches production. His approach to application-specific AI evaluations draws on earlier work in concurrent systems and autonomous driving, where stronger model scores alone cannot establish whether a product behaves safely in the real world.

Goyal studied at the University of California, Berkeley, from 2013 to 2017 and coauthored libcuckoo, an open-source C++ hash-table library supporting concurrent readers and writers. He later worked on self-driving technology at Nuro before joining Braintrust as a founding engineer. There, his engineering interests include semistructured and multimodal data infrastructure, interactive visualization, and nondeterministic AI systems. His contributions to Autoevals include evaluation tracing, SDK integration, packaging, and error handling.

• Evaluate real-world behavior, not isolated model scores. Autonomous vehicles must navigate traffic, avoid pedestrians, and follow road rules; similarly, AI applications require measurements tied to their actual tasks and consequential failure modes.
• Use evaluations as an experimentation laboratory. Offline evaluations let engineers compare models, prompts, and application changes before deployment, reducing the cost and risk of learning exclusively from production.
• Build a production-to-evaluation data flywheel. Applying consistent metrics to offline experiments and production traces helps teams identify real examples for future test cases and connect evaluation with logging, prompt experimentation, and AI observability.
• Make evaluation infrastructure inspectable. Open-source tooling such as Autoevals makes scoring, tracing, SDK integration, and error handling concrete engineering concerns rather than opaque quality claims.

At AI Engineer World’s Fair 2025, Goyal framed evaluations as practical infrastructure for shipping AI systems with greater confidence and improving them through evidence from actual use.